# 2026-06-27: Four Model Onboarding + Benchmarks

## Models Added

| # | HF Repo | File | Size | Target Dir |
|---|---------|------|-----:|------------|
| 1 | `mradermacher/StrikeGPT-VL-8B-GGUF` | `StrikeGPT-VL-8B.Q6_K.gguf` | 6.3 GB | `/models/downloads/StrikeGPT-VL-8B-GGUF/` |
| 2 | `TheBloke/WhiteRabbitNeo-13B-GGUF` | `whiterabbitneo-13b.Q4_K_M.gguf` | 7.4 GB | `/models/downloads/WhiteRabbitNeo-13B-GGUF/` |
| 3 | `empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF` | `Qwythos-9B-Claude-Mythos-5-1M-MTP-Q6_K.gguf` | 7.1 GB | `/models/downloads/Qwythos-9B-Claude-Mythos-5-1M-MTP-GGUF/` |
| 4 | `DavidAU/OpenAi-GPT-oss-20b-abliterated-uncensored-NEO-Imatrix-gguf` | `OpenAI-20B-NEOPlus-Uncensored-Q8_0.gguf` | 21 GB | `/models/downloads/OpenAi-GPT-oss-20b-abliterated-uncensored-NEO-Imatrix-gguf/` |

## Benchmarks (RTX 2080 Ti 11 GB, llama-server-sm75, b9743)

| Model | Tech | TG | PP | VRAM | Config Notes |
|-------|------|-----|-----|------|-------------|
| StrikeGPT-VL-8B Q6 | 8B dense | **69.7 t/s** | 369/65/65 | 8,724 MiB | `-ngl 99`, `-c 32768`, fits entirely on GPU |
| WhiteRabbitNeo-13B Q4_K_M | 13B dense | **17.2 t/s** | 75/16/17 | 10,154 MiB | `-ngl 28`, `-c 16384` — partial GPU offload, 13B Q4 won't fit fully |
| Qwythos-9B MTP Q6 | 9B dense + MTP | **111.0 t/s** | 493/350/350 | 7,854 MiB | `-ngl 99`, `--spec-type draft-mtp --spec-draft-n-max 4` — fastest model in fleet |
| OpenAI-20B-NEOPlus Q8 | 20B MoE Q8 | **18.2 t/s** | 35/36/38 | 1,576 MiB | `--n-cpu-moe 32`, `--no-kv-offload` — mostly CPU offloaded |

## Key Findings

### StrikeGPT-VL-8B
- Vision-language model but LLM-only GGUF (vision tower stripped)
- Q6 at 8B = ~5.5 bits per weight, 6.3 GB on disk
- Fits fully on 11 GB at `-ngl 99` with 32K context
- Good mid-range coder at ~70 t/s

### WhiteRabbitNeo-13B
- Based on Llama 2 13B architecture (dense, 40 layers)
- Q4_K_M = 7.4 GB on disk, but needs 13.6 GB GPU buffer for full offload
- `-ngl 28` fits in ~10.1 GB VRAM → 17.2 t/s
- Could potentially push `-ngl 30-32` depending on KV cache headroom
- Security-focused/red-teaming model, not general purpose

### Qwythos-9B MTP Q6
- Fastest model in the fleet at **111 t/s** — beats even StrikeGPT-8B by 60%
- Has MTP heads for self-speculative decoding
- Fully on-GPU with 7.9 GB VRAM
- Based on Qwen architecture, 1M context natively (limited to 32K for benchmark)
- Likely a merge/recipe model (Qwythos = Qwen + Mythos blend)

### OpenAI-20B-NEOPlus-Uncensored Q8
- Abliterated/uncensored variant of GPT-OSS-20B
- MoE architecture (20B total, ~3B active) — uses `--n-cpu-moe`
- Q8 at 21 GB on disk → mostly CPU offloaded (only 1.6 GB VRAM)
- ~18 t/s is CPU bandwidth limited — comparable to Ornith-35B Q8 was at 18.9 t/s
- `--n-cpu-moe` 32 vs 28 made no significant difference (VRAM was identical = 1576 MiB both ways) — the auto-fitter may be overriding

## Download Notes
- Used `hf download` CLI (not deprecated `huggingface-cli`)
- All downloads parallel, each in its own background process with `notify_on_complete=true`
- StrikeGPT (6.3 GB) fastest, completed in ~90s
- NEOPlus (21 GB) slowest, took ~5 min
- Cache cleaned via `rm -rf <dir>/.cache/` after each download